[AMD] Add MiniMax-M2.5 nightly perf benchmarks for MI30x and MI35x (#21524)

2026-04-03 01:01:03 -07:00
parent 7431db7392
commit d07d0a15ce
4 changed files with 338 additions and 4 deletions
@@ -685,7 +685,7 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

-  # 8-GPU MiniMax-M2.5 (Accuracy) ROCm 7.2
+  # 8-GPU MiniMax-M2.5 (Accuracy + Performance combined) ROCm 7.2
  nightly-8-gpu-minimax-m25-rocm720:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-minimax-m25-rocm720,'))
    runs-on: linux-mi325-8gpu-sglang
@@ -716,6 +716,18 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

+      - name: Performance Test ROCm 7.2 (8-GPU MiniMax-M2.5)
+        timeout-minutes: 120
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-minimax-m25 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
  # ============================================== MI30x ROCm 7.2 Diffusion Tests ==============================================
  # 1-GPU Z-Image-Turbo (Diffusion T2I) ROCm 7.2
  nightly-1-gpu-zimage-turbo-rocm720:
@@ -1306,7 +1318,7 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

-  # MI35x 8-GPU MiniMax-M2.5 (Accuracy) ROCm 7.2
+  # MI35x 8-GPU MiniMax-M2.5 (Accuracy + Performance combined) ROCm 7.2
  nightly-8-gpu-mi35x-minimax-m25-rocm720:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-minimax-m25-rocm720,'))
    runs-on: linux-mi35x-gpu-8
@@ -1339,6 +1351,18 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

+      - name: Performance Test MI35x ROCm 7.2 (8-GPU MiniMax-M2.5)
+        timeout-minutes: 120
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-minimax-m25 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
  # MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP) ROCm 7.2
  nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720,'))
@@ -687,7 +687,7 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

-  # 8-GPU MiniMax-M2.5 (Accuracy)
+  # 8-GPU MiniMax-M2.5 (Accuracy + Performance combined)
  nightly-8-gpu-minimax-m25:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-minimax-m25,'))
    runs-on: linux-mi325-8gpu-sglang
@@ -718,6 +718,18 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

+      - name: Performance Test (8-GPU MiniMax-M2.5)
+        timeout-minutes: 120
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-minimax-m25 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
  # ============================================== MI30x Diffusion Tests ==============================================
  # 1-GPU Z-Image-Turbo (Diffusion T2I)
  nightly-1-gpu-zimage-turbo:
@@ -1278,7 +1290,7 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

-  # MI35x 8-GPU MiniMax-M2.5 (Accuracy)
+  # MI35x 8-GPU MiniMax-M2.5 (Accuracy + Performance combined)
  nightly-8-gpu-mi35x-minimax-m25:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-minimax-m25,'))
    runs-on: linux-mi35x-gpu-8
@@ -1311,6 +1323,18 @@ jobs:
          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
          exit ${TEST_EXIT_CODE:-0}

+      - name: Performance Test MI35x (8-GPU MiniMax-M2.5)
+        timeout-minutes: 120
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-minimax-m25 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
  # MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP)
  nightly-perf-8-gpu-mi35x-deepseek-v32-mtp:
    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-8-gpu-mi35x-deepseek-v32-mtp,'))
@@ -0,0 +1,140 @@
+"""Nightly performance benchmark for MiniMax-M2.5 on MI325/MI300X (8-GPU).
+
+This test benchmarks MiniMax-M2.5 with TP=8 + EP=8 configuration.
+
+The model path can be configured via MINIMAX_M25_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-minimax-m25 suite
+
+Example usage:
+    python -m pytest test_minimax_m25_perf_amd.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+register_amd_ci(est_time=5400, suite="nightly-perf-8-gpu-minimax-m25", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI325")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+MINIMAX_M25_MODEL_PATH = os.environ.get(
+    "MINIMAX_M25_MODEL_PATH", "MiniMaxAI/MiniMax-M2.5"
+)
+PROFILE_DIR = "performance_profiles_minimax_m25"
+
+
+class TestNightlyMiniMaxM25Performance(unittest.TestCase):
+    """Nightly performance benchmark for MiniMax-M2.5 on MI325/MI300X.
+
+    Tests MiniMax-M2.5 with TP=8 + EP=8 configuration.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "minimax-m25-tp8-ep8",
+            "model_path": MINIMAX_M25_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--ep-size",
+                "8",
+                "--attention-backend",
+                "aiter",
+                "--mem-fraction-static",
+                "0.85",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
+            ],
+            "env_vars": {
+                "SGLANG_USE_AITER": "1",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_minimax_m25(self):
+        """Run benchmark for MiniMax-M2.5."""
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+                enable_profile=False,
+                timeout=5400,
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for MiniMax-M2.5")
+        finally:
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,146 @@
+"""MI35x Nightly performance benchmark for MiniMax-M2.5 (8-GPU).
+
+This test benchmarks MiniMax-M2.5 with TP=8 + EP=8 configuration on MI35x.
+
+The model path can be configured via MINIMAX_M25_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-mi35x-minimax-m25 suite
+
+Example usage:
+    python -m pytest test_minimax_m25_perf_mi35x.py -v
+"""
+
+import os
+
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+register_amd_ci(
+    est_time=5400, suite="nightly-perf-8-gpu-mi35x-minimax-m25", nightly=True
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+MINIMAX_M25_MODEL_PATH = os.environ.get(
+    "MINIMAX_M25_MODEL_PATH", "MiniMaxAI/MiniMax-M2.5"
+)
+PROFILE_DIR = "performance_profiles_minimax_m25_mi35x"
+
+
+class TestNightlyMiniMaxM25PerformanceMI35x(unittest.TestCase):
+    """MI35x Nightly performance benchmark for MiniMax-M2.5.
+
+    Tests MiniMax-M2.5 with TP=8 + EP=8 configuration.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "minimax-m25-tp8-ep8",
+            "model_path": MINIMAX_M25_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--ep-size",
+                "8",
+                "--attention-backend",
+                "aiter",
+                "--mem-fraction-static",
+                "0.85",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
+            ],
+            "env_vars": {
+                "SGLANG_USE_AITER": "1",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_minimax_m25(self):
+        """Run benchmark for MiniMax-M2.5."""
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+                enable_profile=False,
+                timeout=5400,
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for MiniMax-M2.5 on MI35x")
+        finally:
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()